|
Microway High-Speed Interconnect Technology
Microway provides a variety of interconnect solutions to meet the
specific demands of your application and data needs. From standard
connectivity to high bandwidth, low latency technologies like
InfiniBand, Microway delivers state-of-the-art solutions in cluster
and storage interconnect.
The contents of the page below are being updated.
Microway FasTree™ Low Latency Modular Switches for InfiniBand
The main benefits of FasTree technology are its distributed modularity, extremely low latency, and
extensibility. These features double the size of the three-hop fat tree fabrics that can be created
with it, while at the same time doubling the size of local low latency domains. The FasTree architecture
enables modular extensible fabrics. Users are now free to experiment with different fabric topologies
that meet their needs and are either non-blocking or slightly blocked. Simulations show that non-blocking
fabrics are often overkill for many MPI applications, increasing latency while at the same time not
providing any real bandwidth benefit. Simply stated, because InfiniBand often provides more bandwidth
than the typical HPC node or application can use, it often does not pay to use truly non-blocked fabrics.
Taking advantage of this extra bandwidth to aggregate nodes into larger local domains provides lower latency
communication between nearby nodes, while at the same time making large fabrics possible. It also provides
system administrators with an opportunity to divide up clusters into groups of 24 nodes, instead of 12,
which keeps users from interfering with each other in clusters that get shared.
Unlike other products which don't work with older products, InfiniBand
products can all inter-operate, making it possible for older slower links
to run with faster and wider links. As a result, the explanation of speed
has to get across the concepts of speed along with the number of lanes.
InfiniBand uses from one to twelve LVDS pairs to transfer information.
Depending on the number of pairs, the switches and HCA's get labelled as
1X, 4X or 12X. Initially only 1X running at 2.5 Gbits per second was
supported. All prior speeds and widths continue to be supported making
it possible to link an older HCA to a new switch. The current wire
speeds per lane are 2.5 Gbs/sec (SDR) and 5.0 Gbs/sec (DDR). QDR links
running at
10.0 Gb/sec will become available in 2007. The actual bandwidth of a
connection is determined by multiplying the number of pairs by the
speeds. The most common speeds today are 4X SDR and 4X DDR, which
deliver 10 and 20 Gb/sec, respecitvely, however it is also possible to
gang three 4X DDR lanes together to produce 12X DDR connection which has
a bandwidth of 60 Gb/sec.
Microway FasTree™ 72 port InfiniBand Switch
The switches also have long-term economic benefits. An interesting characteristic of InfiniBand
switches is that different speeds and widths work together. Therefore, current SDR/DDR 4X models
will be usable as data aggregation devices when QDR technology becomes available in coming years.
The switches were also designed from the ground up to be modular. As a result, the user doesn't
have to purchase a 96 or 144 port switch to build a 32, 48 or 64 port cluster! In addition, the
cabling problem that results when all of the cables in a cluster come together at a single point
can be eliminated, by simply distributing the switches with the cluster nodes. The FasTree
architecture also results in switch fabrics that are extensible and can be used in basic research
in network design. As a user's needs grow, he can change the topology to match the cluster resources
currently owned, adding switches and modifying the fabric layout on the fly, to accommodate
increasingly larger cluster sizes.
FasTree Data Sheet
Click Here for our whitepaper entitled Low Latency Modular Switches for InfiniBand.
Click Here for Visualization Measurement and Improvement of Linux Clusters.
TriCom™ DDR/QDR InfiniBand HCA with Remote Monitoring and Serial Console
Microway's TriCom PCI-E card combines an HCA with the two other special resources needed to
effectively manage HPC nodes - a BMC and a "switchless" serial port console interface.
The HCA portion of the card uses Mellanox InfiniHost III Mem Free and ConnectX silicon. TriCom cards are available running at SDR, DDR and QDR speeds, and TriCom-X cards also offer 10G ethernet capability.
The BMC is an upgraded version of Microway's NodeWatch card, and can monitor up to 13 Fan Tach lines and three voltages and includes five temperature probes. It can control motherboard reset and power on/off, and even operates when the power to a node is turned off. It communicates over a single daisy chained ethernet cable that supports both a half duplex RS-485 control channel and a full duplex RS-422 serial channel. MCMS, Microway's full-featured web-enabled remote cluster control software package is included.
Microway TriCom-X QDR InfiniBand HCA
The serial console interface converts the node's internal COM2 serial port into RS-422, enabling communication
with a master up to 4000 feet away while sharing the same RS-422 lanes with the rest of the cluster. Up to 256
nodes are supported per segment, running at speeds up to
230K baud. The serial console interface is provided by the web
enabled Microway Cluster Management Software.
TriCom Data Sheet
InfiniScope™ InfiniBand Diagnostic Software
InfiniScope displays the traffic of an InfiniBand fabric in real time by collecting data from
the performance registers on all host and switch ports. InfiniScope is compatible with all InfiniBand products
utilizing mellanox silicon. Data can be read up to 100 times per
second without impacting performance of user programs.
The traffic data is shown on a network graph generated by the InfiniBand subnet manager. It
shows both the topology of the network and the host-to-host paths as calculated by the subnet
manager. DDR and SDR connections are marked, and links with 1X lane width are unmistakably shown with
flashing yellow and black lines.
InfiniScope Screen shot with 12-node Cluster
As traffic flows through the fabric, InfiniScope continuously shows the bandwidth transmitted from each
host and switch port using small color coded squares. The performance of each port is indicated by both
the size and color of the squares. In addition, a strip chart shows the recent traffic pattern for one
user-selected port, for all the hosts combined, or for all the ports on one switch. The chart data being
displayed may be easily selected by moving the mouse over the port of interest. Combining Microway
InfiniScope with MPI Link-Checker provides a powerful cluster and application performance diagnosis toolbox.
Request commercial release and information pricing
InfiniScope White Paper (pdf)
InfiniScope Product Page
|
